MW-MADDPG: a meta-learning based decision-making method for collaborative UAV swarm.

Area of Science:

• Robotics
• Artificial Intelligence
• Aerospace Engineering

Background:

• Unmanned Aerial Vehicles (UAVs) are increasingly used due to cost-effectiveness and safety.
• Multi-agent deep reinforcement learning shows promise for UAV swarm coordination.
• Existing methods struggle with complex mission environments and time constraints.

Purpose of the Study:

• To propose a meta-learning based multi-agent deep reinforcement learning approach for UAV swarm cooperative decision-making.
• To address challenges posed by multivariate mission environments and time constraints.
• To enhance the performance and adaptability of UAV swarms.

Main Methods:

• An improved MAML-based multi-agent deep deterministic policy gradient (MADDPG) algorithm is presented.
• An unbiased initialization network is achieved through automatic weighting of meta-learning trajectories.
• A Reward-TD prioritized experience replay technique is introduced, using immediate reward and TD-error.

Main Results:

• The proposed approach demonstrates effective task accomplishment in new scenarios.
• Significant improvements in task success rate, average reward, and robustness were observed.
• The method outperforms existing approaches in UAV swarm cooperative decision-making.

Conclusions:

• The meta-learning based approach offers a viable solution for UAV swarm cooperative decision-making.
• The enhanced MADDPG algorithm and Reward-TD prioritization improve algorithm resilience and sample utilization.
• This research advances the capabilities of autonomous UAV swarms in dynamic environments.